[Quant][Kernel] Remove GPTQ Group/Dynamic Activation Ordering (#54809)

Signed-off-by: Roderick-Wu <roderickwu2003@gmail.com>
Signed-off-by: Roderick Wu <roderickwu2003@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Roderick Wu
2026-09-08 12:45:12 -07:00
committed by GitHub
co-authored by mgoin
parent 1a522b6949
commit 2c9d68f80b
99 changed files with 1182 additions and 3917 deletions
@@ -75,14 +75,14 @@ QUANT_CONFIGS = [
"b_type": "kU4B8",
"thread_configs": THREAD_CONFIGS,
"thread_m_blocks": THREAD_M_BLOCKS,
"group_blocks": [-1, 0, 2, 4, 8],
"group_blocks": [-1, 2, 4, 8],
},
# GPTQ-INT8
{
"b_type": "kU8B128",
"thread_configs": THREAD_CONFIGS,
"thread_m_blocks": THREAD_M_BLOCKS,
"group_blocks": [-1, 0, 2, 4, 8],
"group_blocks": [-1, 2, 4, 8],
},
# FP8
{
@@ -11,11 +11,9 @@
namespace marlin {
template <int const num_threads, int const num_bits, bool const has_perm,
bool is_a_8bit>
template <int const num_threads, int const num_bits, bool is_a_8bit>
__global__ void gptq_marlin_repack_kernel(
uint32_t const* __restrict__ b_q_weight_ptr,
uint32_t const* __restrict__ perm_ptr, uint32_t* __restrict__ out_ptr,
uint32_t const* __restrict__ b_q_weight_ptr, uint32_t* __restrict__ out_ptr,
int size_k, int size_n) {
constexpr int pack_factor = 32 / num_bits;
@@ -44,31 +42,14 @@ __global__ void gptq_marlin_repack_kernel(
extern __shared__ int4 sh[];
constexpr int perm_size = target_tile_k_size / 4;
int4* sh_perm_ptr = sh;
int4* sh_pipe_ptr = sh_perm_ptr;
if constexpr (has_perm) {
sh_pipe_ptr += perm_size;
}
int4* sh_pipe_ptr = sh;
constexpr int tile_ints = target_tile_k_size / pack_factor;
constexpr int stage_n_threads = target_tile_n_size / 4;
constexpr int stage_k_threads = has_perm ? target_tile_k_size : tile_ints;
constexpr int stage_k_threads = tile_ints;
constexpr int stage_size = stage_k_threads * stage_n_threads;
auto load_perm_to_shared = [&](int k_tile_id) {
int first_k_int4 = (k_tile_id * target_tile_k_size) / 4;
int4 const* perm_int4_ptr = reinterpret_cast<int4 const*>(perm_ptr);
if (threadIdx.x < perm_size) {
sh_perm_ptr[threadIdx.x] = perm_int4_ptr[first_k_int4 + threadIdx.x];
}
__syncthreads();
};
auto fetch_to_shared = [&](int pipe, int k_tile_id, int n_tile_id) {
if (n_tile_id >= n_tiles) {
cp_async_fence();
@@ -79,36 +60,17 @@ __global__ void gptq_marlin_repack_kernel(
int4* sh_ptr = sh_pipe_ptr + stage_size * pipe;
if constexpr (has_perm) {
if (threadIdx.x < stage_size) {
auto k_id = threadIdx.x / stage_n_threads;
auto n_id = threadIdx.x % stage_n_threads;
if (threadIdx.x < stage_size) {
auto k_id = threadIdx.x / stage_n_threads;
auto n_id = threadIdx.x % stage_n_threads;
uint32_t const* sh_perm_int_ptr =
reinterpret_cast<uint32_t const*>(sh_perm_ptr);
int first_k = k_tile_id * target_tile_k_size;
int first_k_packed = first_k / pack_factor;
int src_k = sh_perm_int_ptr[k_id];
int src_k_packed = src_k / pack_factor;
cp_async4(
&sh_ptr[k_id * stage_n_threads + n_id],
reinterpret_cast<int4 const*>(&(
b_q_weight_ptr[src_k_packed * size_n + first_n + (n_id * 4)])));
}
} else {
if (threadIdx.x < stage_size) {
auto k_id = threadIdx.x / stage_n_threads;
auto n_id = threadIdx.x % stage_n_threads;
int first_k = k_tile_id * target_tile_k_size;
int first_k_packed = first_k / pack_factor;
cp_async4(&sh_ptr[k_id * stage_n_threads + n_id],
reinterpret_cast<int4 const*>(
&(b_q_weight_ptr[(first_k_packed + k_id) * size_n +
first_n + (n_id * 4)])));
}
cp_async4(&sh_ptr[k_id * stage_n_threads + n_id],
reinterpret_cast<int4 const*>(
&(b_q_weight_ptr[(first_k_packed + k_id) * size_n +
first_n + (n_id * 4)])));
}
cp_async_fence();
@@ -139,56 +101,34 @@ __global__ void gptq_marlin_repack_kernel(
int4* sh_stage_ptr = sh_pipe_ptr + stage_size * pipe;
uint32_t* sh_stage_int_ptr = reinterpret_cast<uint32_t*>(sh_stage_ptr);
uint32_t* sh_perm_int_ptr = reinterpret_cast<uint32_t*>(sh_perm_ptr);
uint32_t vals[8];
if constexpr (has_perm) {
static_assert(!is_a_8bit);
for (int i = 0; i < 4; i++) {
int k_idx = tc_row + tc_offsets[i];
uint32_t src_k = sh_perm_int_ptr[k_idx];
uint32_t src_k_pos = src_k % pack_factor;
uint32_t b1_val = sh_stage_int_ptr[k_idx * sh_stride + cur_n];
uint32_t b1_cur_val = (b1_val >> (src_k_pos * num_bits)) & mask;
uint32_t b2_val = sh_stage_int_ptr[k_idx * sh_stride + cur_n + 8];
uint32_t b2_cur_val = (b2_val >> (src_k_pos * num_bits)) & mask;
vals[i] = b1_cur_val;
vals[4 + i] = b2_cur_val;
}
} else {
uint32_t b1_vals[tile_ints];
uint32_t b2_vals[tile_ints];
uint32_t b1_vals[tile_ints];
uint32_t b2_vals[tile_ints];
#pragma unroll
for (int i = 0; i < tile_ints; i++) {
if constexpr (is_a_8bit) {
b1_vals[i] =
sh_stage_int_ptr[cur_n + sh_stride * i + (warp_id % 2) * 8];
} else {
b1_vals[i] = sh_stage_int_ptr[cur_n + sh_stride * i];
b2_vals[i] = sh_stage_int_ptr[cur_n + 8 + sh_stride * i];
}
for (int i = 0; i < tile_ints; i++) {
if constexpr (is_a_8bit) {
b1_vals[i] =
sh_stage_int_ptr[cur_n + sh_stride * i + (warp_id % 2) * 8];
} else {
b1_vals[i] = sh_stage_int_ptr[cur_n + sh_stride * i];
b2_vals[i] = sh_stage_int_ptr[cur_n + 8 + sh_stride * i];
}
}
#pragma unroll
for (int i = 0; i < 4; i++) {
int cur_elem = tc_row + (is_a_8bit ? i : tc_offsets[i]);
int cur_int = cur_elem / pack_factor;
int cur_pos = cur_elem % pack_factor;
for (int i = 0; i < 4; i++) {
int cur_elem = tc_row + (is_a_8bit ? i : tc_offsets[i]);
int cur_int = cur_elem / pack_factor;
int cur_pos = cur_elem % pack_factor;
vals[i] = (b1_vals[cur_int] >> (cur_pos * num_bits)) & mask;
if constexpr (is_a_8bit)
vals[4 + i] =
(b1_vals[cur_int + tile_ints / 2] >> (cur_pos * num_bits)) & mask;
else
vals[4 + i] = (b2_vals[cur_int] >> (cur_pos * num_bits)) & mask;
}
vals[i] = (b1_vals[cur_int] >> (cur_pos * num_bits)) & mask;
if constexpr (is_a_8bit)
vals[4 + i] =
(b1_vals[cur_int + tile_ints / 2] >> (cur_pos * num_bits)) & mask;
else
vals[4 + i] = (b2_vals[cur_int] >> (cur_pos * num_bits)) & mask;
}
constexpr int tile_size =
@@ -248,10 +188,6 @@ __global__ void gptq_marlin_repack_kernel(
for (int k_tile_id = start_k_tile; k_tile_id < finish_k_tile; k_tile_id++) {
int n_tile_id = 0;
if constexpr (has_perm) {
load_perm_to_shared(k_tile_id);
}
start_pipes(k_tile_id, n_tile_id);
while (n_tile_id < n_tiles) {
@@ -269,21 +205,19 @@ __global__ void gptq_marlin_repack_kernel(
} // namespace marlin
#define CALL_IF(NUM_BITS, HAS_PERM, IS_A_8BIT) \
else if (num_bits == NUM_BITS && has_perm == HAS_PERM && \
is_a_8bit == IS_A_8BIT) { \
#define CALL_IF(NUM_BITS, IS_A_8BIT) \
else if (num_bits == NUM_BITS && is_a_8bit == IS_A_8BIT) { \
cudaFuncSetAttribute( \
marlin::gptq_marlin_repack_kernel<marlin::repack_threads, NUM_BITS, \
HAS_PERM, IS_A_8BIT>, \
IS_A_8BIT>, \
cudaFuncAttributeMaxDynamicSharedMemorySize, max_shared_mem); \
marlin::gptq_marlin_repack_kernel<marlin::repack_threads, NUM_BITS, \
HAS_PERM, IS_A_8BIT> \
IS_A_8BIT> \
<<<blocks, marlin::repack_threads, max_shared_mem, stream>>>( \
b_q_weight_ptr, perm_ptr, out_ptr, size_k, size_n); \
b_q_weight_ptr, out_ptr, size_k, size_n); \
}
torch::stable::Tensor gptq_marlin_repack(torch::stable::Tensor& b_q_weight,
torch::stable::Tensor& perm,
int64_t size_k, int64_t size_n,
int64_t num_bits, bool is_a_8bit) {
// Verify compatibility with marlin tile of 16x64
@@ -311,11 +245,6 @@ torch::stable::Tensor gptq_marlin_repack(torch::stable::Tensor& b_q_weight,
b_q_weight.scalar_type() == torch::headeronly::ScalarType::Int,
"b_q_weight type is not kInt");
STD_TORCH_CHECK(perm.is_cuda(), "perm is not on GPU");
STD_TORCH_CHECK(perm.is_contiguous(), "perm is not contiguous");
STD_TORCH_CHECK(perm.scalar_type() == torch::headeronly::ScalarType::Int,
"perm type is not at::kInt");
const int32_t device_index = b_q_weight.get_device_index();
torch::stable::accelerator::DeviceGuard device_guard(device_index);
const cudaStream_t stream = get_current_cuda_stream(device_index);
@@ -325,14 +254,9 @@ torch::stable::Tensor gptq_marlin_repack(torch::stable::Tensor& b_q_weight,
{size_k / marlin::tile_size, size_n * marlin::tile_size / pack_factor},
b_q_weight.scalar_type(), std::nullopt, b_q_weight.device());
// Detect if there is act_order
bool has_perm = perm.size(0) != 0;
// Get ptrs
uint32_t const* b_q_weight_ptr =
reinterpret_cast<uint32_t const*>(b_q_weight.const_data_ptr());
uint32_t const* perm_ptr =
reinterpret_cast<uint32_t const*>(perm.const_data_ptr());
uint32_t* out_ptr = reinterpret_cast<uint32_t*>(out.mutable_data_ptr());
int blocks;
@@ -345,17 +269,15 @@ torch::stable::Tensor gptq_marlin_repack(torch::stable::Tensor& b_q_weight,
if (false) {
}
CALL_IF(4, false, false)
CALL_IF(4, true, false)
CALL_IF(8, false, false)
CALL_IF(8, true, false)
CALL_IF(4, false)
CALL_IF(8, false)
CALL_IF(4, false, true)
CALL_IF(8, false, true)
CALL_IF(4, true)
CALL_IF(8, true)
else {
STD_TORCH_CHECK(false, "Unsupported repack config: num_bits = ", num_bits,
", has_perm = ", has_perm, ", is_a_8bit = ", is_a_8bit);
", is_a_8bit = ", is_a_8bit);
}
return out;
@@ -7,17 +7,16 @@
#include "marlin_dtypes.cuh"
#include "core/scalar_type.hpp"
#define MARLIN_KERNEL_PARAMS \
const int4 *__restrict__ A, const int4 *__restrict__ B, \
int4 *__restrict__ C, int4 *__restrict__ C_tmp, \
const int4 *__restrict__ b_bias_ptr, \
const float *__restrict__ a_scales_ptr, \
const int4 *__restrict__ scales_ptr, \
const float *__restrict__ global_scale_ptr, \
const int4 *__restrict__ zp_ptr, const int *__restrict__ g_idx, \
int num_groups, int prob_m, int prob_n, int prob_k, int lda, int *locks, \
bool has_bias, bool use_atomic_add, bool use_fp32_reduce, \
int max_shared_mem
#define MARLIN_KERNEL_PARAMS \
const int4 *__restrict__ A, const int4 *__restrict__ B, \
int4 *__restrict__ C, int4 *__restrict__ C_tmp, \
const int4 *__restrict__ b_bias_ptr, \
const float *__restrict__ a_scales_ptr, \
const int4 *__restrict__ scales_ptr, \
const float *__restrict__ global_scale_ptr, \
const int4 *__restrict__ zp_ptr, int prob_m, int prob_n, int prob_k, \
int lda, int *locks, bool has_bias, bool use_atomic_add, \
bool use_fp32_reduce, int max_shared_mem
namespace MARLIN_NAMESPACE_NAME {
template <const vllm::ScalarTypeId a_type_id, // A ScalarType id
@@ -47,11 +47,6 @@ using MarlinFuncPtr = void (*)(MARLIN_KERNEL_PARAMS);
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
__global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
int const* __restrict__ perm_int_ptr,
int4* __restrict__ out_int4_ptr, int size_m,
int size_k, int lda, int block_rows) {}
} // namespace marlin
torch::stable::Tensor marlin_gemm(
@@ -62,11 +57,9 @@ torch::stable::Tensor marlin_gemm(
std::optional<torch::stable::Tensor> const& a_scales_or_none,
std::optional<torch::stable::Tensor> const& global_scale_or_none,
std::optional<torch::stable::Tensor> const& b_zeros_or_none,
std::optional<torch::stable::Tensor> const& g_idx_or_none,
std::optional<torch::stable::Tensor> const& perm_or_none,
torch::stable::Tensor& workspace, vllm::ScalarTypeId const& b_type_id,
int64_t size_m, int64_t size_n, int64_t size_k, bool is_k_full,
bool use_atomic_add, bool use_fp32_reduce, bool is_zp_float) {
int64_t size_m, int64_t size_n, int64_t size_k, bool use_atomic_add,
bool use_fp32_reduce, bool is_zp_float) {
STD_TORCH_CHECK_NOT_IMPLEMENTED(false,
"marlin_gemm(..) requires CUDA_ARCH >= 7.5");
return torch::stable::empty({1, 1});
@@ -74,62 +67,6 @@ torch::stable::Tensor marlin_gemm(
#else
// For a given "a" of size [M,K] performs a permutation of the K columns based
// on the given "perm" indices.
__global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
int const* __restrict__ perm_int_ptr,
int4* __restrict__ out_int4_ptr, int size_m,
int size_k, int lda, int block_rows) {
auto start_row = block_rows * blockIdx.x;
int finish_row = start_row + block_rows;
if (finish_row > size_m) {
finish_row = size_m;
}
int cur_block_rows = finish_row - start_row;
int input_row_stride = lda * sizeof(half) / 16;
int output_row_stride = size_k * sizeof(half) / 16;
auto permute_row = [&](int row) {
int iters = size_k / default_threads;
int rest = size_k % default_threads;
int input_offset = row * input_row_stride;
int output_offset = row * output_row_stride;
half const* a_row_half =
reinterpret_cast<half const*>(a_int4_ptr + input_offset);
half* out_half = reinterpret_cast<half*>(out_int4_ptr + output_offset);
int base_k = 0;
for (int i = 0; i < iters; i++) {
auto cur_k = base_k + threadIdx.x;
int src_pos = perm_int_ptr[cur_k];
out_half[cur_k] = a_row_half[src_pos];
base_k += default_threads;
}
if (rest) {
if (threadIdx.x < rest) {
auto cur_k = base_k + threadIdx.x;
int src_pos = perm_int_ptr[cur_k];
out_half[cur_k] = a_row_half[src_pos];
}
}
};
for (int i = 0; i < cur_block_rows; i++) {
int cur_row = start_row + i;
if (cur_row < size_m) {
permute_row(cur_row);
}
}
}
typedef struct {
int thread_k;
int thread_n;
@@ -159,9 +96,7 @@ typedef struct {
int get_scales_cache_size(thread_config_t const& th_config, int prob_m,
int prob_n, int prob_k, int num_bits, int group_size,
bool has_act_order, bool is_k_full, int stages) {
bool cache_scales_chunk = has_act_order && !is_k_full;
int stages) {
int tb_n = th_config.thread_n;
int tb_k = th_config.thread_k;
@@ -169,29 +104,18 @@ int get_scales_cache_size(thread_config_t const& th_config, int prob_m,
int tb_groups;
if (group_size == -1) {
tb_groups = 1;
} else if (group_size == 0) {
tb_groups = div_ceil(tb_k, 32); // Worst case is 32 group size
} else {
tb_groups = div_ceil(tb_k, group_size);
}
if (cache_scales_chunk) {
int load_groups =
tb_groups * stages * 2; // Chunk size is 2x pipeline over dim K
load_groups = max(load_groups, 32); // We load at least 32 scale groups
return load_groups * tb_n * 2;
} else {
int tb_scales = tb_groups * tb_n * 2;
return tb_scales * stages;
}
int tb_scales = tb_groups * tb_n * 2;
return tb_scales * stages;
}
int get_kernel_cache_size(thread_config_t const& th_config, int thread_m_blocks,
int prob_m, int prob_n, int prob_k, int num_bits,
int group_size, bool has_act_order, bool is_k_full,
int has_zp, bool is_zp_float, bool is_a_8bit,
int stages) {
int group_size, int has_zp, bool is_zp_float,
bool is_a_8bit, int stages) {
int pack_factor = 32 / num_bits;
// Get B size
@@ -206,10 +130,8 @@ int get_kernel_cache_size(thread_config_t const& th_config, int thread_m_blocks,
(sh_b_size > sh_red_size ? sh_red_size : sh_b_size) + sh_bias_size;
tmp_size = max(max(sh_b_size, sh_red_size), tmp_size);
int sh_s_size =
get_scales_cache_size(th_config, prob_m, prob_n, prob_k, num_bits,
group_size, has_act_order, is_k_full, stages);
int sh_g_idx_size = has_act_order && !is_k_full ? stages * tb_k / 4 : 0;
int sh_s_size = get_scales_cache_size(th_config, prob_m, prob_n, prob_k,
num_bits, group_size, stages);
int sh_zp_size = 0;
if (has_zp) {
if (is_zp_float)
@@ -220,17 +142,15 @@ int get_kernel_cache_size(thread_config_t const& th_config, int thread_m_blocks,
sh_zp_size = sh_s_size / 2;
}
int total_size =
tmp_size + sh_a_size + sh_s_size + sh_zp_size + sh_g_idx_size;
int total_size = tmp_size + sh_a_size + sh_s_size + sh_zp_size;
return total_size;
}
bool is_valid_config(thread_config_t const& th_config, int thread_m_blocks,
int prob_m, int prob_n, int prob_k, int num_bits,
int group_size, bool has_act_order, bool is_k_full,
int has_zp, bool is_zp_float, bool is_a_8bit, int stages,
int max_shared_mem) {
int group_size, int has_zp, bool is_zp_float,
bool is_a_8bit, int stages, int max_shared_mem) {
// Sanity
if (th_config.thread_k == -1 || th_config.thread_n == -1 ||
th_config.num_threads == -1) {
@@ -255,16 +175,18 @@ bool is_valid_config(thread_config_t const& th_config, int thread_m_blocks,
// Check that pipeline fits into cache
int cache_size = get_kernel_cache_size(
th_config, thread_m_blocks, prob_m, prob_n, prob_k, num_bits, group_size,
has_act_order, is_k_full, has_zp, is_zp_float, is_a_8bit, stages);
has_zp, is_zp_float, is_a_8bit, stages);
return cache_size <= max_shared_mem;
}
MarlinFuncPtr get_marlin_kernel(
const vllm::ScalarType a_type, const vllm::ScalarType b_type,
const vllm::ScalarType c_type, const vllm::ScalarType s_type,
int thread_m_blocks, int thread_n_blocks, int thread_k_blocks,
bool m_block_size_8, bool has_act_order, bool has_zp, int group_blocks,
int threads, bool is_zp_float, int stages) {
MarlinFuncPtr get_marlin_kernel(const vllm::ScalarType a_type,
const vllm::ScalarType b_type,
const vllm::ScalarType c_type,
const vllm::ScalarType s_type,
int thread_m_blocks, int thread_n_blocks,
int thread_k_blocks, bool m_block_size_8,
bool has_zp, int group_blocks, int threads,
bool is_zp_float, int stages) {
int num_bits = b_type.size_bits();
auto kernel = MarlinDefault;
@@ -277,9 +199,8 @@ exec_config_t determine_exec_config(
const vllm::ScalarType& a_type, const vllm::ScalarType& b_type,
const vllm::ScalarType& c_type, const vllm::ScalarType& s_type, int prob_m,
int prob_n, int prob_k, int thread_m_blocks, bool m_block_size_8,
int num_bits, int group_size, bool has_act_order, bool is_k_full,
bool has_zp, bool is_zp_float, int is_a_8bit, int stages,
int max_shared_mem, int sms) {
int num_bits, int group_size, bool has_zp, bool is_zp_float, int is_a_8bit,
int stages, int max_shared_mem, int sms) {
exec_config_t exec_cfg = exec_config_t{1, thread_config_t{-1, -1, -1}};
thread_config_t* thread_configs = thread_m_blocks > 1
? large_batch_thread_configs
@@ -293,27 +214,21 @@ exec_config_t determine_exec_config(
thread_config_t th_config = thread_configs[i];
if (!is_valid_config(th_config, thread_m_blocks, prob_m, prob_n, prob_k,
num_bits, group_size, has_act_order, is_k_full, has_zp,
is_zp_float, is_a_8bit, stages,
max_shared_mem - 512)) {
num_bits, group_size, has_zp, is_zp_float, is_a_8bit,
stages, max_shared_mem - 512)) {
continue;
}
int cache_size = get_kernel_cache_size(th_config, thread_m_blocks, prob_m,
prob_n, prob_k, num_bits, group_size,
has_act_order, is_k_full, has_zp,
is_zp_float, is_a_8bit, stages);
int cache_size = get_kernel_cache_size(
th_config, thread_m_blocks, prob_m, prob_n, prob_k, num_bits,
group_size, has_zp, is_zp_float, is_a_8bit, stages);
int group_blocks = 0;
if (!has_act_order) {
group_blocks = group_size == -1 ? -1 : group_size / 16;
}
int group_blocks = group_size == -1 ? -1 : group_size / 16;
auto kernel =
get_marlin_kernel(a_type, b_type, c_type, s_type, thread_m_blocks,
th_config.thread_n / 16, th_config.thread_k / 16,
m_block_size_8, has_act_order, has_zp, group_blocks,
th_config.num_threads, is_zp_float, stages);
auto kernel = get_marlin_kernel(
a_type, b_type, c_type, s_type, thread_m_blocks,
th_config.thread_n / 16, th_config.thread_k / 16, m_block_size_8,
has_zp, group_blocks, th_config.num_threads, is_zp_float, stages);
if (kernel == MarlinDefault) continue;
@@ -324,38 +239,25 @@ exec_config_t determine_exec_config(
}
void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
void* a_s, void* b_s, void* g_s, void* zp, void* g_idx,
void* perm, void* a_tmp, int prob_m, int prob_n, int prob_k,
int lda, void* workspace, vllm::ScalarType const& a_type,
vllm::ScalarType const& b_type, vllm::ScalarType const& c_type,
vllm::ScalarType const& s_type, bool has_bias,
bool has_act_order, bool is_k_full, bool has_zp, int num_groups,
int group_size, int dev, cudaStream_t stream, int thread_k_init,
void* a_s, void* b_s, void* g_s, void* zp, int prob_m,
int prob_n, int prob_k, int lda, void* workspace,
vllm::ScalarType const& a_type, vllm::ScalarType const& b_type,
vllm::ScalarType const& c_type, vllm::ScalarType const& s_type,
bool has_bias, bool has_zp, int num_groups, int group_size,
int dev, cudaStream_t stream, int thread_k_init,
int thread_n_init, int sms, bool use_atomic_add,
bool use_fp32_reduce, bool is_zp_float) {
bool is_a_8bit = a_type.size_bits() == 8;
STD_TORCH_CHECK(prob_m > 0 && prob_n > 0 && prob_k > 0, "Invalid MNK = [",
prob_m, ", ", prob_n, ", ", prob_k, "]");
int group_blocks = 0;
if (has_act_order) {
if (is_k_full) {
STD_TORCH_CHECK(group_size != -1);
group_blocks = group_size / 16;
STD_TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
" is not divisible by group_blocks = ", group_blocks);
} else {
STD_TORCH_CHECK(group_size == 0);
group_blocks = 0;
}
int group_blocks;
if (group_size == -1) {
group_blocks = -1;
} else {
if (group_size == -1) {
group_blocks = -1;
} else {
group_blocks = group_size / 16;
STD_TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
" is not divisible by group_blocks = ", group_blocks);
}
group_blocks = group_size / 16;
STD_TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
" is not divisible by group_blocks = ", group_blocks);
}
int num_bits = b_type.size_bits();
@@ -370,28 +272,8 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
const float* g_s_ptr = (const float*)g_s;
const int4* zp_ptr = (const int4*)zp;
const int* g_idx_ptr = (const int*)g_idx;
const int* perm_ptr = (const int*)perm;
int4* a_tmp_ptr = (int4*)a_tmp;
int* locks = (int*)workspace;
if (has_act_order) {
// Permute A columns
int block_rows = div_ceil(prob_m, sms);
// avoid ">>>" being formatted to "> > >"
// clang-format off
permute_cols_kernel<<<sms, default_threads, 0, stream>>>(
A_ptr, perm_ptr, a_tmp_ptr, prob_m, prob_k, lda, block_rows);
// clang-format on
A_ptr = a_tmp_ptr;
lda = prob_k;
// If we have a full K, then we can run the non-act-order version of Marlin
// (since the weight rows are reordered by increasing group ids, and by
// having a full K, we have full original groups)
if (is_k_full) has_act_order = false;
}
int max_shared_mem = 0;
cudaDeviceGetAttribute(&max_shared_mem,
cudaDevAttrMaxSharedMemoryPerBlockOptin, dev);
@@ -451,18 +333,17 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
// Auto config
exec_cfg = determine_exec_config(
a_type, b_type, c_type, s_type, prob_m_split, prob_n, prob_k,
thread_m_blocks, m_block_size_8, num_bits, group_size, has_act_order,
is_k_full, has_zp, is_zp_float, is_a_8bit, stages, max_shared_mem,
sms);
thread_m_blocks, m_block_size_8, num_bits, group_size, has_zp,
is_zp_float, is_a_8bit, stages, max_shared_mem, sms);
thread_tfg = exec_cfg.tb_cfg;
if (thread_tfg.thread_n != -1) {
if (prob_n / thread_tfg.thread_n *
div_ceil(prob_m_split, thread_m_blocks * 16) * 4 <=
sms) {
if (is_valid_config({128, 64, 128}, thread_m_blocks, prob_m_split,
prob_n, prob_k, num_bits, group_size,
has_act_order, is_k_full, has_zp, is_zp_float,
is_a_8bit, stages, max_shared_mem_new)) {
prob_n, prob_k, num_bits, group_size, has_zp,
is_zp_float, is_a_8bit, stages,
max_shared_mem_new)) {
thread_tfg = {128, 64, 128};
exec_cfg = {1, thread_tfg};
}
@@ -487,30 +368,27 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
STD_TORCH_CHECK(
is_valid_config(thread_tfg, thread_m_blocks, prob_m_split, prob_n,
prob_k, num_bits, group_size, has_act_order, is_k_full,
has_zp, is_zp_float, is_a_8bit, stages,
max_shared_mem_new),
prob_k, num_bits, group_size, has_zp, is_zp_float,
is_a_8bit, stages, max_shared_mem_new),
"Invalid thread config: thread_m_blocks = ", thread_m_blocks,
", thread_k = ", thread_tfg.thread_k,
", thread_n = ", thread_tfg.thread_n,
", num_threads = ", thread_tfg.num_threads, " for MKN = [", prob_m,
", ", prob_k, ", ", prob_n, "] and num_bits = ", num_bits,
", prob_m_split = ", prob_m_split, ", group_size = ", group_size,
", has_act_order = ", has_act_order, ", is_k_full = ", is_k_full,
", has_zp = ", has_zp, ", is_zp_float = ", is_zp_float,
", stages = ", stages, ", max_shared_mem_new = ", max_shared_mem_new);
auto kernel = get_marlin_kernel(
a_type, b_type, c_type, s_type, thread_m_blocks, thread_n_blocks,
thread_k_blocks, m_block_size_8, has_act_order, has_zp, group_blocks,
num_threads, is_zp_float, stages);
thread_k_blocks, m_block_size_8, has_zp, group_blocks, num_threads,
is_zp_float, stages);
if (kernel == MarlinDefault) {
STD_TORCH_CHECK(
false, "Unsupported shapes: MNK = [", prob_m, ", ", prob_n, ", ",
prob_k, "]", ", has_act_order = ", has_act_order,
", num_groups = ", num_groups, ", group_size = ", group_size,
", prob_m_split = ", prob_m_split,
prob_k, "]", ", num_groups = ", num_groups,
", group_size = ", group_size, ", prob_m_split = ", prob_m_split,
", thread_m_blocks = ", thread_m_blocks,
", thread_n_blocks = ", thread_n_blocks,
", thread_k_blocks = ", thread_k_blocks,
@@ -527,7 +405,6 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
// clang-format off
kernel<<<blocks, num_threads, max_shared_mem_new, stream>>>(
A_ptr, B_ptr, C_ptr, C_tmp_ptr, bias_ptr, a_s_ptr, b_s_ptr, g_s_ptr, zp_ptr,
g_idx_ptr, num_groups,
prob_m_split, prob_n, prob_k, lda, locks, has_bias, part_use_atomic_add,
use_fp32_reduce, max_shared_mem_new);
// clang-format on
@@ -550,11 +427,9 @@ torch::stable::Tensor marlin_gemm(
std::optional<torch::stable::Tensor> const& a_scales_or_none,
std::optional<torch::stable::Tensor> const& global_scale_or_none,
std::optional<torch::stable::Tensor> const& b_zeros_or_none,
std::optional<torch::stable::Tensor> const& g_idx_or_none,
std::optional<torch::stable::Tensor> const& perm_or_none,
torch::stable::Tensor& workspace, vllm::ScalarTypeId const& b_type_id,
int64_t size_m, int64_t size_n, int64_t size_k, bool is_k_full,
bool use_atomic_add, bool use_fp32_reduce, bool is_zp_float) {
int64_t size_m, int64_t size_n, int64_t size_k, bool use_atomic_add,
bool use_fp32_reduce, bool is_zp_float) {
vllm::ScalarTypeId a_type_id, c_type_id, s_type_id;
auto c_scalar_type = a.scalar_type();
@@ -722,7 +597,7 @@ torch::stable::Tensor marlin_gemm(
std::nullopt, device);
}
// Detect groupsize and act_order
// Detect group size.
int num_groups = -1;
int group_size = -1;
@@ -733,51 +608,13 @@ torch::stable::Tensor marlin_gemm(
" is not size_n = ", size_n);
num_groups = b_scales.size(0);
torch::stable::Tensor g_idx, perm, a_tmp;
if (g_idx_or_none.has_value() && perm_or_none.has_value()) {
g_idx = g_idx_or_none.value();
perm = perm_or_none.value();
STD_TORCH_CHECK(g_idx.device().is_cuda(), "g_idx is not on GPU");
STD_TORCH_CHECK(g_idx.is_contiguous(), "g_idx is not contiguous");
STD_TORCH_CHECK(perm.device().is_cuda(), "perm is not on GPU");
STD_TORCH_CHECK(perm.is_contiguous(), "perm is not contiguous");
// Verify g_idx and perm
STD_TORCH_CHECK((g_idx.size(-1) == 0 && perm.size(-1) == 0) ||
(g_idx.size(-1) == size_k && perm.size(-1) == size_k),
"Unexpected g_idx.size(-1) = ", g_idx.size(-1),
" and perm.size(-1) = ", perm.size(-1),
", where size_k = ", size_k);
if (num_groups > 1) {
STD_TORCH_CHECK(
size_k % num_groups == 0, "size_k = ", size_k,
", is not divisible by b_scales.size(0) = ", b_scales.size(0));
group_size = size_k / num_groups;
} else {
g_idx = torch::stable::empty({0}, c_scalar_type, std::nullopt, device);
perm = torch::stable::empty({0}, c_scalar_type, std::nullopt, device);
a_tmp = torch::stable::empty({0}, c_scalar_type, std::nullopt, device);
}
bool has_act_order = g_idx.size(-1) > 0 && perm.size(-1) > 0;
if (has_act_order) {
a_tmp = torch::stable::empty({size_m, size_k}, c_scalar_type, std::nullopt,
device);
if (is_k_full) {
STD_TORCH_CHECK(num_groups > 1, "For act_order, num_groups must be > 1");
STD_TORCH_CHECK(size_k % num_groups == 0, "size_k = ", size_k,
", is not divisible by num_groups = ", num_groups);
group_size = size_k / num_groups;
} else {
group_size = 0;
}
} else {
a_tmp = torch::stable::empty({0}, c_scalar_type, std::nullopt, device);
if (num_groups > 1) {
STD_TORCH_CHECK(
size_k % num_groups == 0, "size_k = ", size_k,
", is not divisible by b_scales.size(0) = ", b_scales.size(0));
group_size = size_k / num_groups;
} else {
group_size = -1;
}
group_size = -1;
}
torch::stable::Tensor global_scale;
@@ -882,11 +719,9 @@ torch::stable::Tensor marlin_gemm(
a.const_data_ptr(), b_q_weight.const_data_ptr(), c.mutable_data_ptr(),
c_tmp.mutable_data_ptr(), b_bias.mutable_data_ptr(),
a_scales.mutable_data_ptr(), b_scales.mutable_data_ptr(),
global_scale.mutable_data_ptr(), b_zeros.mutable_data_ptr(),
g_idx.mutable_data_ptr(), perm.mutable_data_ptr(),
a_tmp.mutable_data_ptr(), size_m, size_n, size_k, a.stride(0),
workspace.mutable_data_ptr(), a_type, b_type, c_type, s_type, has_bias,
has_act_order, is_k_full, has_zp, num_groups, group_size, device_index,
global_scale.mutable_data_ptr(), b_zeros.mutable_data_ptr(), size_m,
size_n, size_k, a.stride(0), workspace.mutable_data_ptr(), a_type, b_type,
c_type, s_type, has_bias, has_zp, num_groups, group_size, device_index,
get_current_cuda_stream(device_index), thread_k, thread_n, sms,
use_atomic_add, use_fp32_reduce, is_zp_float);
@@ -51,10 +51,9 @@ template <typename scalar_t, // compute dtype, half or nv_float16
// only works when thread_m_blocks == 1
const int stages, // number of stages for the async global->shared
// fetch pipeline
const bool has_act_order, // whether act_order is enabled
const int group_blocks, // number of consecutive 16x16 blocks
// with a separate quantization scale
const bool is_zp_float // is zero point of float16 type?
const int group_blocks, // number of consecutive 16x16 blocks
// with a separate quantization scale
const bool is_zp_float // is zero point of float16 type?
>
__global__ void Marlin(
const int4* __restrict__ A, // fp16 input matrix of shape mxk
@@ -63,11 +62,9 @@ __global__ void Marlin(
int4* __restrict__ C_tmp, // fp32 tmp output buffer (for reduce)
const int4* __restrict__ scales_ptr, // fp16 quantization scales of shape
// (k/groupsize)xn
const int* __restrict__ g_idx, // int32 group indices of shape k
int num_groups, // number of scale groups per output channel
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int* locks, // extra global storage for barrier synchronization
bool use_fp32_reduce // whether to use fp32 global reduce
) {}
@@ -140,29 +137,6 @@ __device__ inline void sub_zp(
frag_b[1] = __hsub2(frag_b[1], zp);
}
// Same as above, but for act_order (each K is multiplied individually)
template <vllm::ScalarTypeId type_id>
__device__ inline void scale4(
typename MarlinScalarType<type_id>::FragB& frag_b,
typename MarlinScalarType<type_id>::FragS& frag_s_1,
typename MarlinScalarType<type_id>::FragS& frag_s_2,
typename MarlinScalarType<type_id>::FragS& frag_s_3,
typename MarlinScalarType<type_id>::FragS& frag_s_4, int i) {
using scalar_t = typename MarlinScalarType<type_id>::scalar_t;
using scalar_t2 = typename MarlinScalarType<type_id>::scalar_t2;
scalar_t2 s_val_1_2;
s_val_1_2.x = reinterpret_cast<scalar_t*>(&frag_s_1)[i];
s_val_1_2.y = reinterpret_cast<scalar_t*>(&frag_s_2)[i];
scalar_t2 s_val_3_4;
s_val_3_4.x = reinterpret_cast<scalar_t*>(&frag_s_3)[i];
s_val_3_4.y = reinterpret_cast<scalar_t*>(&frag_s_4)[i];
frag_b[0] = __hmul2(frag_b[0], s_val_1_2);
frag_b[1] = __hmul2(frag_b[1], s_val_3_4);
}
// Given 2 floats multiply by 2 scales (halves)
template <vllm::ScalarTypeId type_id>
__device__ inline void scale_float(
@@ -256,14 +230,11 @@ __global__ void Marlin(
// 4bit packed zero-points of shape
// (k/groupsize, n/pack_factor)
const int4* __restrict__ zp_ptr,
// int32 group indices of shape k
const int* __restrict__ g_idx,
int num_groups, // number of scale groups per output channel
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int lda, // A.stride(0), equal to prob_k is A is contiguous
int* locks, // extra global storage for barrier synchronization
int prob_m, // batch dimension m
int prob_n, // output dimension n
int prob_k, // reduction dimension k
int lda, // A.stride(0), equal to prob_k is A is contiguous
int* locks, // extra global storage for barrier synchronization
bool has_bias,
bool use_atomic_add, // whether to use atomic add to reduce
bool use_fp32_reduce, // whether to use fp32 global reduce
@@ -358,7 +329,8 @@ __global__ void Marlin(
global_scale_f32 = global_scale_ptr[0];
}
constexpr bool has_act_order = group_blocks == 0;
static_assert(group_blocks != 0,
"runtime activation ordering is not supported");
constexpr int m_block_size = m_block_size_8 ? 8 : (16 * thread_m_blocks);
extern __shared__ int4 sh[];
@@ -391,7 +363,7 @@ __global__ void Marlin(
int iters = div_ceil(k_tiles * part2_mn_tiles, gridDim.x);
if constexpr (!has_act_order && group_blocks != -1) {
if constexpr (group_blocks != -1) {
if (group_blocks >= thread_k_blocks) {
// Ensure that the number of tiles in each stripe is a multiple of the
// groupsize; this avoids an annoying special case where a stripe starts
@@ -558,27 +530,17 @@ __global__ void Marlin(
b_sh_stride * thread_k_blocks / (is_a_8bit ? 2 : 1);
constexpr int b_sh_wr_iters = b_sh_stage / b_sh_wr_delta;
// Scale sizes/strides without act_order
// Scale sizes/strides
int s_gl_stride = prob_n / (is_8bit_scale ? 16 : 8);
constexpr int s_sh_stride = 16 * thread_n_blocks / (is_8bit_scale ? 16 : 8);
constexpr int s_tb_groups =
!has_act_order && group_blocks != -1 && group_blocks < thread_k_blocks
group_blocks != -1 && group_blocks < thread_k_blocks
? thread_k_blocks / group_blocks
: 1;
constexpr int s_sh_stage = s_tb_groups * s_sh_stride;
int s_gl_rd_delta = s_gl_stride;
// Scale size/strides with act_order
constexpr int tb_k = 16 * thread_k_blocks;
constexpr int g_idx_stage = has_act_order ? (tb_k * sizeof(int)) / 16 : 0;
// constexpr int act_s_row_stride = 1;
// int act_s_col_stride = act_s_row_stride * num_groups;
constexpr int act_s_max_num_groups = 32;
int act_s_col_stride = 1;
int act_s_col_warp_stride = act_s_col_stride * 8;
constexpr int tb_n_warps = thread_n_blocks / (is_a_8bit ? 2 : 4);
int act_s_col_tb_stride = act_s_col_warp_stride * tb_n_warps;
// Zero-points sizes/strides
int zp_gl_stride = is_zp_float ? prob_n / 8 : (prob_n / pack_factor) / 4;
@@ -615,25 +577,16 @@ __global__ void Marlin(
auto b_sh_rd = threadIdx.x * b_thread_vecs;
b_sh_rd += b_sh_rd / b_sh_stride * (b_sh_stride * (b_sh_wr_iters - 1));
// For act_order
int slice_k_start = tb_k * slice_row;
int slice_k_finish = slice_k_start + tb_k * slice_iters;
int slice_k_start_shared_fetch = slice_k_start;
int slice_n_offset = act_s_col_tb_stride * slice_col;
// No act_order
int s_gl_rd;
if constexpr (!has_act_order) {
if constexpr (group_blocks == -1) {
s_gl_rd = s_sh_stride * slice_col + threadIdx.x;
} else if constexpr (group_blocks >= thread_k_blocks) {
s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
s_sh_stride * slice_col + threadIdx.x;
} else {
s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks +
threadIdx.x / s_sh_stride) +
s_sh_stride * slice_col + threadIdx.x % s_sh_stride;
}
if constexpr (group_blocks == -1) {
s_gl_rd = s_sh_stride * slice_col + threadIdx.x;
} else if constexpr (group_blocks >= thread_k_blocks) {
s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
s_sh_stride * slice_col + threadIdx.x;
} else {
s_gl_rd = s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks +
threadIdx.x / s_sh_stride) +
s_sh_stride * slice_col + threadIdx.x % s_sh_stride;
}
auto s_sh_wr = threadIdx.x;
bool s_sh_wr_pred = threadIdx.x < s_sh_stage;
@@ -757,10 +710,8 @@ __global__ void Marlin(
: (sh_size_b_red_min + sh_bias_size);
int4* sh_bias = sh_new + sh_size_b_red_min;
int4* sh_g_idx = sh_new + sh_b_red_bias_size;
int4* sh_zp = sh_g_idx + (stages * g_idx_stage);
constexpr int sh_s_size = has_act_order ? (act_s_max_num_groups * s_sh_stride)
: (stages * s_sh_stage);
int4* sh_zp = sh_new + sh_b_red_bias_size;
constexpr int sh_s_size = stages * s_sh_stage;
int4* sh_s = sh_zp + (stages * zp_sh_stage);
int4* sh_a = sh_s + sh_s_size;
@@ -769,9 +720,8 @@ __global__ void Marlin(
I4 frag_b_quant[2][b_thread_vecs];
FragC frag_c[thread_m_blocks][is_a_8bit ? 2 : 4][2];
FragC frag_c_tmp[thread_m_blocks][is_a_8bit ? 2 : 4][2];
FragS frag_s[2][4]; // No act-order
FragS frag_s[2][4];
FragS frag_bias[2][4];
FragS act_frag_s[2][4][4]; // For act-order
int frag_qzp[2][num_ints_per_thread]; // Zero-points
FragZP frag_zp; // Zero-points in fp16
FragZP frag_zpf[2]; // Zero-points in fp16 in HQQ
@@ -801,42 +751,6 @@ __global__ void Marlin(
reinterpret_cast<float*>(frag_c)[i] = 0;
};
int sh_first_group_id = -1;
int sh_num_groups = -1;
auto fetch_act_order_scales_to_shared = [&](bool is_async, int first_group_id,
int last_group_id) {
sh_first_group_id = first_group_id;
sh_num_groups = last_group_id - first_group_id + 1;
if (sh_num_groups > act_s_max_num_groups) {
sh_num_groups = act_s_max_num_groups;
}
if (sh_first_group_id + sh_num_groups > num_groups) {
sh_num_groups = num_groups - sh_first_group_id;
}
int row_offset = first_group_id * s_gl_stride;
if (is_async) {
for (int i = 0; i < sh_num_groups; i++) {
if (threadIdx.x < s_sh_stride) {
cp_async4_pred(&sh_s[(i * s_sh_stride) + threadIdx.x],
&scales_ptr[row_offset + (i * s_gl_stride) +
slice_n_offset + threadIdx.x]);
}
}
} else {
for (int i = 0; i < sh_num_groups; i++) {
if (threadIdx.x < s_sh_stride) {
sh_s[(i * s_sh_stride) + threadIdx.x] =
scales_ptr[row_offset + (i * s_gl_stride) + slice_n_offset +
threadIdx.x];
}
}
}
};
// Asynchronously fetch the next A, B and s tile from global to the next
// shared memory pipeline location.
auto fetch_to_shared = [&](int pipe, int a_off, bool pred = true) {
@@ -862,44 +776,27 @@ __global__ void Marlin(
b_gl_rd += b_gl_rd_delta_o;
if constexpr (has_act_order) {
// Fetch g_idx thread-block portion
int full_pipe = a_off;
int cur_k = slice_k_start_shared_fetch + tb_k * full_pipe;
if (cur_k < prob_k && cur_k < slice_k_finish) {
int4* sh_g_idx_stage = sh_g_idx + g_idx_stage * pipe;
if constexpr (group_blocks != -1) {
int4* sh_s_stage = sh_s + s_sh_stage * pipe;
int4 const* cur_g_idx_stage_ptr =
reinterpret_cast<int4 const*>(&g_idx[cur_k]);
if (threadIdx.x < g_idx_stage) {
cp_async4_pred(&sh_g_idx_stage[threadIdx.x],
&cur_g_idx_stage_ptr[threadIdx.x]);
// Only fetch scales if this tile starts a new group
if (pipe % div_ceil(group_blocks, thread_k_blocks) == 0) {
if (s_sh_wr_pred) {
cp_async4(&sh_s_stage[s_sh_wr], &scales_ptr[s_gl_rd]);
}
s_gl_rd += s_gl_rd_delta * s_tb_groups;
}
} else {
if constexpr (group_blocks != -1) {
int4* sh_s_stage = sh_s + s_sh_stage * pipe;
}
// Only fetch scales if this tile starts a new group
if (pipe % div_ceil(group_blocks, thread_k_blocks) == 0) {
if (s_sh_wr_pred) {
cp_async4(&sh_s_stage[s_sh_wr], &scales_ptr[s_gl_rd]);
}
s_gl_rd += s_gl_rd_delta * s_tb_groups;
}
}
if constexpr (has_zp && group_blocks != -1) {
int4* sh_zp_stage = sh_zp + zp_sh_stage * pipe;
if constexpr (has_zp && group_blocks != -1) {
int4* sh_zp_stage = sh_zp + zp_sh_stage * pipe;
// Only fetch zero points if this tile starts a new group
if (pipe % div_ceil(group_blocks, thread_k_blocks) == 0) {
if (zp_sh_wr_pred) {
cp_async4(&sh_zp_stage[zp_sh_wr], &zp_ptr[zp_gl_rd]);
}
zp_gl_rd += zp_gl_rd_delta * zp_tb_groups;
// Only fetch zero points if this tile starts a new group
if (pipe % div_ceil(group_blocks, thread_k_blocks) == 0) {
if (zp_sh_wr_pred) {
cp_async4(&sh_zp_stage[zp_sh_wr], &zp_ptr[zp_gl_rd]);
}
zp_gl_rd += zp_gl_rd_delta * zp_tb_groups;
}
}
}
@@ -947,150 +844,57 @@ __global__ void Marlin(
}
};
bool is_same_group[stages];
int same_group_id[stages];
auto init_same_group = [&](int pipe) {
if constexpr (!has_act_order) {
return;
}
int4* sh_g_idx_stage = sh_g_idx + g_idx_stage * pipe;
int* sh_g_idx_int_ptr = reinterpret_cast<int*>(sh_g_idx_stage);
int group_id_1 = sh_g_idx_int_ptr[0];
int group_id_2 = sh_g_idx_int_ptr[tb_k - 1];
is_same_group[pipe] = group_id_1 == group_id_2;
same_group_id[pipe] = group_id_1;
};
auto fetch_scales_to_registers = [&](int k, int full_pipe) {
int pipe = full_pipe % stages;
using IT1 = typename std::conditional_t<is_a_8bit, int2, int4>;
using IT0 = typename std::conditional_t<is_a_8bit, int, int2>;
constexpr int group_blocks2 = div_ceil(group_blocks, is_a_8bit ? 2 : 1);
if constexpr (!has_act_order) {
// No act-order case
if constexpr (group_blocks == -1) {
// load only when starting a new slice
if (k == 0 && full_pipe == 0 && dequant_skip_flop) {
reinterpret_cast<int4*>(&frag_s)[0] = sh_s[s_sh_rd];
reinterpret_cast<int4*>(&frag_s)[1] = sh_s[s_sh_rd + 4];
}
} else if constexpr (group_blocks != -1) {
if constexpr (group_blocks >= thread_k_blocks) {
constexpr int g = group_blocks / thread_k_blocks;
if (pipe % g == 0) {
if (k % b_sh_wr_iters == 0) {
int4* sh_s_stage = sh_s + s_sh_stage * (g * (pipe / g));
reinterpret_cast<int4*>(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd];
} else {
reinterpret_cast<int4*>(&frag_s[1])[0] =
reinterpret_cast<int4*>(&frag_s[0])[0];
}
}
} else if (group_blocks2 < b_sh_wr_iters || k % b_sh_wr_iters == 0) {
auto warp_id = threadIdx.x / 32;
int warp_row = warp_id / tb_n_warps;
int k_blocks = b_sh_wr_iters * warp_row + k % b_sh_wr_iters;
int cur_group_id = k_blocks / group_blocks2;
int4* sh_s_stage = sh_s + s_sh_stage * pipe;
if constexpr (!is_8bit_scale) {
reinterpret_cast<int4*>(&frag_s[k % 2])[0] =
sh_s_stage[s_sh_rd + cur_group_id * s_sh_stride];
} else {
reinterpret_cast<int2*>(&frag_s[k % 2])[0] =
reinterpret_cast<int2*>(
sh_s_stage)[s_sh_rd + cur_group_id * (2 * s_sh_stride)];
}
} else if (group_blocks >= b_sh_wr_iters) {
if constexpr (!is_8bit_scale) {
reinterpret_cast<int4*>(&frag_s[1])[0] =
reinterpret_cast<int4*>(&frag_s[0])[0];
} else {
reinterpret_cast<int2*>(&frag_s[1])[0] =
reinterpret_cast<int2*>(&frag_s[0])[0];
}
if constexpr (group_blocks == -1) {
// Load only when starting a new slice.
if (k == 0 && full_pipe == 0 && dequant_skip_flop) {
reinterpret_cast<int4*>(&frag_s)[0] = sh_s[s_sh_rd];
reinterpret_cast<int4*>(&frag_s)[1] = sh_s[s_sh_rd + 4];
}
} else if constexpr (group_blocks >= thread_k_blocks) {
constexpr int g = group_blocks / thread_k_blocks;
if (pipe % g == 0) {
if (k % b_sh_wr_iters == 0) {
int4* sh_s_stage = sh_s + s_sh_stage * (g * (pipe / g));
reinterpret_cast<int4*>(&frag_s[k % 2])[0] = sh_s_stage[s_sh_rd];
} else {
reinterpret_cast<int4*>(&frag_s[1])[0] =
reinterpret_cast<int4*>(&frag_s[0])[0];
}
}
} else if (group_blocks2 < b_sh_wr_iters || k % b_sh_wr_iters == 0) {
auto warp_id = threadIdx.x / 32;
int warp_row = warp_id / tb_n_warps;
return;
}
int k_blocks = b_sh_wr_iters * warp_row + k % b_sh_wr_iters;
int cur_group_id = k_blocks / group_blocks2;
// Act-order case
int4* sh_s_stage = sh_s + s_sh_stage * pipe;
// Determine K of the "current" thread-block
int cur_k = slice_k_start + tb_k * full_pipe;
if (cur_k >= prob_k || cur_k >= slice_k_finish) {
return;
}
// Reset (to current thread-block) since we read g_idx portion from the
// shared memory
cur_k = 0;
// Progress to current iteration
cur_k += k % b_sh_wr_iters;
// Determine "position" inside the thread-block (based on warp and
// thread-id)
auto warp_id = threadIdx.x / 32;
int warp_row = warp_id / tb_n_warps;
int warp_col = warp_id % tb_n_warps;
cur_k += warp_row * 16 * b_sh_wr_iters;
auto th_id = threadIdx.x % 32;
cur_k += (th_id % 4) * 2; // Due to tensor-core layout for fp16 B matrix
int s_col_shift =
/*slice_n_offset +*/ (act_s_col_warp_stride * warp_col) +
(th_id / 4) * act_s_col_stride;
if (is_same_group[pipe]) {
if (k % 2 == 0) {
*(reinterpret_cast<int4*>(&(act_frag_s[k % 2][0][0]))) =
sh_s[(same_group_id[pipe] - sh_first_group_id) * s_sh_stride +
s_col_shift];
if constexpr (!is_8bit_scale) {
reinterpret_cast<int4*>(&frag_s[k % 2])[0] =
sh_s_stage[s_sh_rd + cur_group_id * s_sh_stride];
} else {
*(reinterpret_cast<int4*>(&(act_frag_s[k % 2][0][0]))) =
*(reinterpret_cast<int4*>(&(act_frag_s[(k - 1) % 2][0][0])));
reinterpret_cast<int2*>(&frag_s[k % 2])[0] = reinterpret_cast<int2*>(
sh_s_stage)[s_sh_rd + cur_group_id * (2 * s_sh_stride)];
}
for (int i = 1; i < 4; i++) {
*(reinterpret_cast<int4*>(&(act_frag_s[k % 2][i][0]))) =
*(reinterpret_cast<int4*>(&(act_frag_s[k % 2][0][0])));
} else if (group_blocks >= b_sh_wr_iters) {
if constexpr (!is_8bit_scale) {
reinterpret_cast<int4*>(&frag_s[1])[0] =
reinterpret_cast<int4*>(&frag_s[0])[0];
} else {
reinterpret_cast<int2*>(&frag_s[1])[0] =
reinterpret_cast<int2*>(&frag_s[0])[0];
}
return;
}
int4* sh_g_idx_stage = sh_g_idx + g_idx_stage * pipe;
int* sh_g_idx_int_ptr = reinterpret_cast<int*>(sh_g_idx_stage);
constexpr int k_frag_offsets[4] = {0, 1, 8,
9}; // Tensor core offsets per thread
#pragma unroll
for (int i = 0; i < 4; i++) {
int actual_k = cur_k + k_frag_offsets[i];
int group_id = sh_g_idx_int_ptr[actual_k];
int rel_group_id = group_id - sh_first_group_id;
*(reinterpret_cast<int4*>(&(act_frag_s[k % 2][i][0]))) =
sh_s[rel_group_id * s_sh_stride + s_col_shift];
}
};
auto fetch_zp_to_registers = [&](int k, int full_pipe) {
// This code does not handle group_blocks == 0,
// which signifies act_order.
// has_zp implies AWQ, which doesn't have act_order,
static_assert(!has_zp || group_blocks != 0);
if constexpr (has_zp && !is_zp_float) {
@@ -1180,8 +984,7 @@ __global__ void Marlin(
constexpr int g =
group_blocks > 0 ? div_ceil(group_blocks, thread_k_blocks) : 1;
const bool is_new_zp =
(group_blocks == 0) ||
((group_blocks > 0) && (group_blocks < b_sh_wr_iters || k == 0)) &&
(group_blocks > 0 && (group_blocks < b_sh_wr_iters || k == 0)) &&
(pipe % g == 0) ||
(group_blocks == -1 && is_first_matmul_in_slice);
if constexpr (has_zp && !is_zp_float) {
@@ -1249,15 +1052,8 @@ __global__ void Marlin(
sub_zp<a_type_id>(frag_b1, frag_zp[j], 1);
}
// Apply scale to frag_b0
if constexpr (has_act_order && !is_a_8bit) {
static_assert(group_blocks != -1);
scale4<a_type_id>(frag_b0, act_frag_s[k2][0][j], act_frag_s[k2][1][j],
act_frag_s[k2][2][j], act_frag_s[k2][3][j], 0);
scale4<a_type_id>(frag_b1, act_frag_s[k2][0][j], act_frag_s[k2][1][j],
act_frag_s[k2][2][j], act_frag_s[k2][3][j], 1);
} else if constexpr (!dequant_skip_flop && has_zp && !is_zp_float &&
group_blocks == -1 && !is_a_8bit) {
if constexpr (!dequant_skip_flop && has_zp && !is_zp_float &&
group_blocks == -1 && !is_a_8bit) {
int idx = (threadIdx.x / 4) % 2;
scalar_t2 s2 = Adtype::nums2num2(
reinterpret_cast<scalar_t*>(&frag_s[j / 2][j % 2 * 2 + 0])[idx],
@@ -1661,7 +1457,7 @@ __global__ void Marlin(
// For per-column quantization we finally apply the scale here (only for
// 4-bit)
if constexpr (!has_act_order && group_blocks == -1 && !is_a_8bit &&
if constexpr (group_blocks == -1 && !is_a_8bit &&
b_type.size_bits() == 4 &&
(has_zp && dequant_skip_flop || !has_zp)) {
c_scalar_t2 tmp_scale = s[0];
@@ -1750,15 +1546,6 @@ __global__ void Marlin(
#pragma unroll
for (int i = 0; i < stages - 1; i++) {
if (has_act_order && i == 0) {
int last_g_idx = slice_k_start + stages * tb_k * 2;
if (last_g_idx >= prob_k) {
last_g_idx = prob_k - 1;
}
fetch_act_order_scales_to_shared(true, g_idx[slice_k_start],
g_idx[last_g_idx]);
}
if constexpr (has_zp && !is_zp_float && group_blocks == -1) {
if (i == 0) {
fetch_col_zp_to_shared();
@@ -1772,14 +1559,10 @@ __global__ void Marlin(
zero_accums();
wait_for_stage();
init_same_group(0);
fetch_to_registers(0, 0);
fetch_scales_to_registers(0, 0);
fetch_zp_to_registers(0, 0);
a_gl_rd += a_gl_rd_delta_o * (stages - 1);
if constexpr (has_act_order) {
slice_k_start_shared_fetch += tb_k * (stages - 1);
}
};
if (slice_iters) {
start_pipes();
@@ -1804,7 +1587,6 @@ __global__ void Marlin(
slice_iters >= stages);
pipe++;
wait_for_stage();
init_same_group(pipe % stages);
}
if constexpr (!is_a_8bit) {
@@ -1822,25 +1604,6 @@ __global__ void Marlin(
a_gl_rd += a_gl_rd_delta_o * stages;
if constexpr (has_act_order) {
slice_k_start += tb_k * stages;
if (slice_k_start < prob_k) {
slice_k_start_shared_fetch += tb_k * stages;
int first_group_id = g_idx[slice_k_start];
int last_g_idx = slice_k_start + stages * tb_k * 2;
if (last_g_idx >= prob_k) {
last_g_idx = prob_k - 1;
}
int last_group_id = g_idx[last_g_idx];
if (last_group_id >= sh_first_group_id + sh_num_groups) {
fetch_act_order_scales_to_shared(false, first_group_id,
last_group_id);
__syncthreads();
}
}
}
// Process results and, if necessary, proceed to the next column slice.
// While this pattern may not be the most readable, other ways of writing
// the loop seemed to noticeably worse performance after compilation.
@@ -1898,7 +1661,7 @@ __global__ void Marlin(
bool last = slice_idx == slice_count - 1;
// For per-column scales, we only fetch them here in the final step before
// write-out
if constexpr (!has_act_order && group_blocks == -1 &&
if constexpr (group_blocks == -1 &&
(has_zp && dequant_skip_flop || !has_zp)) {
if (b_type.size_bits() == 8 || (last || use_atomic_add) || is_a_8bit) {
if (s_sh_wr_pred) {
@@ -1917,7 +1680,7 @@ __global__ void Marlin(
cp_async_fence();
}
if constexpr (!has_act_order && group_blocks == -1 &&
if constexpr (group_blocks == -1 &&
(has_zp && dequant_skip_flop || !has_zp || is_a_8bit)) {
if constexpr (is_a_8bit) {
cp_async_wait<0>();
@@ -1948,7 +1711,7 @@ __global__ void Marlin(
// For 8-bit channelwise, we apply the scale before the global reduction
// that converts the fp32 results to fp16 (so that we avoid possible
// overflow in fp16)
if constexpr (!has_act_order && group_blocks == -1 && is_a_8bit) {
if constexpr (group_blocks == -1 && is_a_8bit) {
#pragma unroll
for (int j = 0; j < 2; j++) {
float2 aa[2];
@@ -1970,8 +1733,7 @@ __global__ void Marlin(
}
}
}
} else if (!has_act_order && group_blocks == -1 &&
b_type.size_bits() == 8 &&
} else if (group_blocks == -1 && b_type.size_bits() == 8 &&
(has_zp && dequant_skip_flop || !has_zp)) {
if (threadIdx.x / 32 < tb_n_warps) {
#pragma unroll
@@ -2042,33 +1804,25 @@ __global__ void Marlin(
b_gl_rd += b_sh_stride * slice_col + b_gl_rd_delta_o * slice_row;
bias_gl_rd = (thread_n_blocks * 16 / 8) * slice_col + threadIdx.x;
// Update slice k/n for scales loading
if constexpr (has_act_order) {
slice_k_start = tb_k * slice_row;
slice_k_finish = slice_k_start + tb_k * slice_iters;
slice_k_start_shared_fetch = slice_k_start;
slice_n_offset = act_s_col_tb_stride * slice_col;
if constexpr (group_blocks == -1) {
s_gl_rd = s_sh_stride * slice_col + threadIdx.x;
zp_gl_rd = zp_sh_stride * slice_col + threadIdx.x;
} else if constexpr (group_blocks >= thread_k_blocks) {
s_gl_rd =
s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
s_sh_stride * slice_col + threadIdx.x;
zp_gl_rd =
zp_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
zp_sh_stride * slice_col + threadIdx.x;
} else {
if constexpr (group_blocks == -1) {
s_gl_rd = s_sh_stride * slice_col + threadIdx.x;
zp_gl_rd = zp_sh_stride * slice_col + threadIdx.x;
} else if constexpr (group_blocks >= thread_k_blocks) {
s_gl_rd =
s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
s_sh_stride * slice_col + threadIdx.x;
zp_gl_rd =
zp_gl_stride * ((thread_k_blocks * slice_row) / group_blocks) +
zp_sh_stride * slice_col + threadIdx.x;
} else {
s_gl_rd =
s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks +
threadIdx.x / s_sh_stride) +
s_sh_stride * slice_col + threadIdx.x % s_sh_stride;
zp_gl_rd =
zp_gl_stride * ((thread_k_blocks * slice_row) / group_blocks +
threadIdx.x / zp_sh_stride) +
zp_sh_stride * slice_col + threadIdx.x % zp_sh_stride;
}
s_gl_rd =
s_gl_stride * ((thread_k_blocks * slice_row) / group_blocks +
threadIdx.x / s_sh_stride) +
s_sh_stride * slice_col + threadIdx.x % s_sh_stride;
zp_gl_rd =
zp_gl_stride * ((thread_k_blocks * slice_row) / group_blocks +
threadIdx.x / zp_sh_stride) +
zp_sh_stride * slice_col + threadIdx.x % zp_sh_stride;
}
start_pipes();
}